CITI has stopped operations in 2014, to co-launch NOVA LINCS THIS SITE IS NOT BEING UPDATED SINCE 2013
citi banner
Home Page FCT/UNL UNL
  Home  \  Graduation Activities  \  Post-Graduation Page Login  
   
banner bottom
File Top
Parallelization using MapReduce over the Cell/BE processor
MSc Post-Graduation
Abstract:

In this dissertation we present the results of an evaluation of an existing MapReduce implementation for the Cell/BE processor. On a first level, we have studied the internal architecture of the runtime system and how it was implemented on a Cell BE architecture. This study provides a detailed analysis of the components of the architecture and how its functionalities were distributed in this design, among the PPE and SPE nodes, as a first step to assist in the interpretation of the experimental evaluation of MapReducecell that was our second level of analysis. We present the results of an experimental evaluation of that MapReduce implementation on the existing Cell BE computer of the CITI/CENTRIA Centers of the Department of Informatics of the Faculty of Science and Technology of the New University of Lisbon. The experimental performance evaluation of MapReduce-cell was focused on a case study of the word count application. An extensive collection of execution tests was conducted, in order to obtain a detailed analysis of all the behavior of the MapReduce execution stages, under the variation of several configuration parameters, in particular the number of SPE, the input data size, and the number of intermediate data partitions considered in the MapReduce execution. The main goal of this study was to assess how this implementation of MapReduce for the Cell would be able to provide acceptable performance and scalability (with respect to the input data size and the number of SPE) for a text mining application. This evaluation allowed to identify the main limitations of the design of MapReduce-cell and to propose alternative solutions to improve the implementation. As an additional stage in our dissertation work, we have evaluated the internal architecture design and its experimental performance behavior, of a MapReduce implementation for a cluster of Cell processors. The methodology for this evaluation was conducted with the same two levels of analysis, namely, a first study of its internal software architecture, and a second experimental evaluation under the word count case study.





Post-Graduation Student / Researcher / Professor:
  • Pedro Afonso ( Departamento de Informática FCT/UNL )

Post-Graduation Supervisor(s):
File Bottom